BoundedGS: Memory-Efficient High-Resolution Training for 3D Gaussian Splatting

BoundedGS is a memory-efficient training method for high-resolution 3D Gaussian Splatting. It processes each view in bounded local regions, separating rasterization and projection backward passes so that cached data can be released after each region contributes its gradients. On five native 5K outdoor scenes, it reduced the mean rendering working set from 9.97 to 1.76 GiB (82.3%) while keeping similar reconstruction quality and cutting training time from 36.7 to 33.9 minutes.